| Intelligent driving is one of the directions of future car development,and environmental perception is the basis of intelligent driving systems.In the environmental perception task,the vehicle needs to obtain environmental information,identify objects,and understand the driving scene through various sensors,so as to provide reference for subsequent decision-making and path planning.The detection of road information is very important for intelligent driving and driving safety.It enables vehicles to effectively judge the drivable area and provide guidance for path planning and safe driving.In the structured urban road scene,the driving environment is complex and the traffic is congested,these problems bring great challenges to the environment perception.According to the above problems,this thesis studies the environmental perception task of autonomous vehicles in road scenes,and proposes an improved method based on the target detection of laser point cloud data,which can effectively solve the occlusion effect of distant vehicle targets in the task of 3D target detection.The problem of detection effect is of great significance to the safety of autonomous driving.The specific research content of this paper is as follows.Aiming at the problem of target detection of autonomous vehicles in road scenes,the feature clustering of point cloud data is carried out through the PCL point cloud processing library,and the characteristics of road vehicles are obtained according to the clustering results.The deep learning method is applied to the processing of 3D point cloud data to solve the problem that the method of manually extracting obstacle features is time-consuming and computationally intensive.Convolutional neural network features are used to extract point cloud features,and the mainstream deep learning detection network frameworks are compared and analyzed.Identify research methods and ideas for improvement.In view of the sparse point cloud data and the problem of missed detection and false detection caused by the occlusion of distant vehicle targets,a multi-head attention mechanism is designed based on the Point Pillars detection network,and a method of grouping similar shape categories is proposed to slow down the multi-category dominant network and improve the target detection effect.To solve the problem of unbalanced samples in the model training data set,a data enhancement operation is introduced to randomly sample vehicle target samples to enhance data training,alleviate the problem of sample imbalance,and increase random noise at the target point cloud to improve noise anti-interference ability;network algorithm model for target detection For complex and time-consuming inference problems,a model compression method is proposed,using Tensor RT to accelerate inference data accuracy calibration,quantize FP32 to INT8,compress model volume,improve inference speed and reduce latency.In the KITTI data set,the evaluation indicators of the improved algorithm model are compared and verified,and the accuracy of target detection is judged by AP.The AP index of3 D detection increased by 7.63%,3.44% and 2.12% respectively at the simple,medium and difficult levels,and the AP index of BEV detection increased by 3.09%,2.57% and 1.12% at the simple,medium and difficult levels,respectively.Prove the effectiveness of the improved object detection algorithm.By building an experimental platform and loading the improved target detection algorithm,the improved target detection algorithm is evaluated in the actual road environment.According to the visualization results,the Multi-Head attention mechanism Point Pillars model improves the detection and recognition rate under the condition of vehicle target occlusion,and the algorithm consumes a single frame.The time is increased by 21 ms,which verifies the effectiveness of the Tensor RT accelerated inference method and meets the real-time requirements.The improved target detection method has the feasibility of being used in unmanned vehicles. |
PDF Full Text Request